config: the dead record parameters go

[ch-wan]

The series

Stacked, each PR based on the one above it. Read them in order; every boundary is
self-sufficient and green on its own.

1. #36896 — the resolution pipeline's dispatcher leaves the record
2. #36972 — the callbacks into the record go to zero
3. #36973 — six more runtime readers ask the bags
4. #36974 — the dead record parameters go ← you are here
5. #36975 — the lazy imports that buy nothing become eager

• #36925 — CI vehicle — runs the whole series against main, not for merge

Motivation

Once a function stops reading the record, the parameter it was handed is dead weight — and a dead
server_args parameter is worse than noise, because the next reader assumes the function consults
the configuration when it does not.

Modifications

A census of the tree finds 256 functions that take a server_args they never read. Most are not
dead, and the distinction matters:

• multimodal_gen's pipeline stages take it as an interface contract — 72 implementations of
  forward use it and 33 do not, so the parameter belongs to the signature, not to the body.
  component_uses is 16 to 29, create_pipeline_stages 13 to 31.
• The same holds for SRTPlatform.apply_server_args_defaults (implemented in other
  distributions), CustomSpecAlgo.handle_server_args, the model-override providers, and the
  StackStrategy.build family.

Ten in srt/ are genuinely dead — free functions called by name, nothing polymorphic about them:

generate_prefill_cuda_graph_batch_sizes   validate_ib_devices
validate_buckets_rule                     handle_hardware_runtime_validation
build_kv_only_stack         build_hybrid_swa_stack
build_hybrid_mamba_stack    build_hybrid_mamba_swa_stack
build_anchor_sidecar_stack  build_deepseek_v4_hicache_stack

The parameter goes from each, and the argument from all eighteen call sites.

The helper that fed a dropped parameter goes too: _validate_ib_devices in the CPU backend test
still built a bare ServerArgs.__new__(ServerArgs) for a call that stopped taking one.

Removing one uncovers the next. test_dead_server_args_parameter_ratchet — added by an earlier
round of this series with a baseline of zero — caught two functions this PR created:
attach_hybrid_minimax_sparse_pool_to_hiradix_cache and its DSA sibling only carried a record to
pass along to the build_*_stack functions above, so they became dead the moment those did. The
removal is iterated to a fixed point now, which is what the ratchet's own docstring asks for.

An eleventh candidate was not dead, and it became #36921 rather than part of this PR:
create_kt_config_from_server_args reached the record through server_args.get_hf_config(), a
method ServerArgs does not have and never had, under a bare except Exception: pass. The call
raised AttributeError every time, num_layers was always None, and the branch that stops the
model's last MoE layer from deferring experts had never fired.

Accuracy Tests

The 62-shape resolution probe is byte-identical to the base commit. The suites that cover the
touched files report the same failure set as main — the five test_fp8_blockwise_linear_backends
failures are a PyTorch interpreter assert present on both sides.

Speed Tests and Profiling

None.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

Pure deletion; the only thing to check is the interface/free-function split above. On this tree
the ratchet's own scan reports 0 hits over 268 files, and the @_register_for override helpers
(_gemma2_gemma3_overrides, _exaone_overrides) still take an unused record because that is their
registration contract — the same reason the polymorphic surfaces are left alone.

---

CI States

Latest PR Test (Base): 🚫 Run #33249912287
Latest PR Test (Extra): ❌ Run #33249912199
Latest PR Test (AMD ROCm 7.2): 🚫 Run #33249912252

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-08-29T06:37:18.886883Z	7425957	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.